MoSim

Official artifact for the IEEE MASCOTS 2026 paper
Accurate Simulation of Distributed Training Jobs with Network Contention Modeling

Yeonho Yoo^†, Hyunho Lee^†, Hyunmok Choi^†, Chuck Yoo^*, and Gyeongsik Yang^*

MoSim is a trace-driven GPU-cluster simulator that models how worker placement and shared network interfaces affect distributed-training job completion times. This repository contains the cluster-level simulator, prepared traces and profiles, physical-testbed results, and scripts that reproduce Tables II–IV of the paper without requiring a GPU.

Artifact scope

The bundled artifact reproduces Tables II, III, and IV. It does not include the ASTRA-sim profile-generation workflow or the source measurements used for Fig. 2, Fig. 4, and Table V.

Quick start

Requirement	Version
Rust	1.78+
Python	3.9+; standard library only for the reproduction scripts
OS	Linux or macOS

From the repository root:

cargo build --release --locked --manifest-path mosim/Cargo.toml
bash scripts/run.sh
bash scripts/analyze.sh

The simulation and analysis normally finish in a few seconds after the Rust build. Generated CSV and Markdown tables are written to result/paper/tab{2,3,4}/out/; result/rawdata.csv contains the combined per-job outputs.

The paper uses the following artifact configurations. Tiresias and Pollux denote the baseline configurations evaluated in the paper; their original implementations are not vendored here.

Method	Single-job profile	Interference model
Tiresias	real-GPU profiling	none
Pollux	real-GPU profiling	fixed 10% networking-time penalty
MoSim	ASTRA-sim	dynamic MoSim model

To rerun experiments after changing parameters or prepared inputs, bypass cached outputs:

FORCE=1 FORCE_PREPARE=1 bash scripts/run.sh

See GUIDE.ipynb for an executable walkthrough. Run python3 simulator-trace-timer-bw.py --help for individual simulator options. Jupyter is only required for the notebook, not for the reproduction scripts.

Repository map

Path	Purpose
simulator-trace-timer-bw.py	User-facing CLI
mosim/src/gpu_job.rs	Job model and bandwidth demand
mosim/src/gpu_cluster.rs	Placement and network-contention models
mosim/src/gpu_scheduler.rs, mosim/src/timer.rs	Scheduling and discrete-event execution
scripts/run.sh, scripts/analyze.py	Paper experiments and Tables II–IV
data/	Prepared traces, profiles, and physical-testbed results

Data provenance

• data/single_job_simulation/ contains precomputed profiles generated using ASTRA-sim. ASTRA-sim itself is not vendored in this repository. When using these profiles, please also cite Won et al., “ASTRA-sim2.0,” ISPASS 2023 (DOI).
• data/trace/ contains modified workload samples derived from the Microsoft Philly trace by Jeon et al. The source dataset is distributed under CC BY 4.0; these files select and transform records into MoSim's trace schema. Please cite “Analysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads”.
• data/profiling/ and data/testbed/ contain measurements collected on the authors' V100 testbed.

Citation

If you use MoSim, please cite the paper. GitHub also exposes the same metadata from CITATION.cff.

@inproceedings{yoo2026accurate,
  author={Yoo, Yeonho and Lee, Hyunho and Choi, Hyunmok and Yoo, Chuck and Yang, Gyeongsik},
  title={Accurate Simulation of Distributed Training Jobs with Network Contention Modeling},
  booktitle={2026 34th IEEE International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems (MASCOTS)},
  year={2026},
  address={Genova, Italy},
  publisher={IEEE}
}

License

The MoSim source code is licensed under the MIT License; see LICENSE. Upstream tools and datasets remain subject to their respective terms.

Acknowledgments

This research was supported by National Research Foundation of Korea (NRF) grant funded by Korea government (MSIT) (RS-2024-00336564), by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by Ministry of Science and ICT (MSIT) (RS-2026-25518394), by ICT Creative Consilience Program through IITP grant funded by MSIT (IITP-2026-RS-2020-II201819), by IITP under the Artificial Intelligence Convergence Innovation Human Resources Development grant funded by Korea government (MSIT) (IITP-2026-RS-2023-00254592), by ANCHOR through Seoul ANCHOR Center funded by MOE and Seoul Metropolitan Government (2026-ANCHOR-01-003-09), and by computing support from Lambda Cloud.

Contact

For artifact, code, or reproduction questions, open an issue. For questions about the paper, contact the corresponding authors, Chuck Yoo and Gyeongsik Yang.